Image Super-resolution Reconstruction Method Based On Image Geometric Structure

In recent years,the application of convolutional neural networks in the field of super-resolution reconstruction has made some breakthroughs,but it still cannot effectively restore geometric structures such as image edges and textures.This paper focuses on how to effectively restore geometric structures such as image edges and textures.The main work is as follows:(1)An image super-resolution reconstruction method combining perceptual edge constraints and multi-scale fusion is proposed.The model of this method includes two stages: the first stage extracts low-resolution image features and upsamples the features to obtain rough high-resolution image features;the second stage network takes the coarse-precision image features as input and passes the feature pyramid module The rough features are refined in a step-by-step codec manner,thereby completing the accurate reconstruction of image edges and textures.In the second stage,considering the difference in the importance of different scale features,the channel attention mechanism is used to capture the channel weights of different scale features,and the weights are used to guide feature fusion.In addition,in order to enhance the network's ability to restore the image edge structure,a perceptual edge constraint based on the image edge geometry structure feature is proposed.Experimental results show that,compared with the current mainstream super-resolution reconstruction method,this method not only achieves a certain improvement in the evaluation index,but also can reconstruct a clearer image edge.(2)In order to describe the edge structure of the image more accurately,a perceptual edge detection method based on context enhanced network is proposed.This method assumes that there is a certain inherent relationship between multi-scale context information,analyzes and models the multi-scale context from the time dimension,uses the idea of recurrent neural network,and uses the recursive neural network branches from top to bottom and bottom to top respectively from two Multi-directional contextual connections are captured in all directions to complete the enhancement of multi-scale contextual features.Comparing this method with the current mainstream edge detection method,the experimental results show that this method can more accurately describe the edge structure of the image.(3)In order to better restore the edge texture geometry of the reconstructed image,an image super-resolution reconstruction method combining perceptual edge constraints and generating an adversarial network is proposed.This method uses the feature extraction network based on the context-enhanced network's perceptual edge detection method to extract the edge structure features,and uses this as a supervision to encourage the generation of network reconstruction to more effectively reconstruct the image edge structure.At the same time,the identification network can extract some difficult to learn.The latent pattern of high-resolution images prompts the generated network to better fit the real sample distribution.Experimental results show that this method can restore the edge texture of the image more effectively than the current mainstream super-resolution reconstruction method,and reconstruct a more realistic image in visual perception.